[UX] Deduplicate sampling parameter startup logs (#32953)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/config/model.py

@@ -1339,10 +1339,9 @@ class ModelConfig:
         Returns:
             A dictionary containing the non-default sampling parameters.
         """
-        if self.generation_config == "vllm":
-            config = {}
-        else:
-            config = self.try_get_generation_config()
+        src = self.generation_config
+
+        config = {} if src == "vllm" else self.try_get_generation_config()
 
         # Overriding with given generation config
         config.update(self.override_generation_config)
@@ -1368,13 +1367,16 @@ class ModelConfig:
         else:
             diff_sampling_param = {}
 
-        if diff_sampling_param:
+        if diff_sampling_param and src != "vllm":
             logger.warning_once(
-                "Default sampling parameters have been overridden by the "
-                "model's Hugging Face generation config recommended from the "
-                "model creator. If this is not intended, please relaunch "
-                "vLLM instance with `--generation-config vllm`."
+                "Default vLLM sampling parameters have been overridden by %s: `%s`. "
+                "If this is not intended, please relaunch vLLM instance "
+                "with `--generation-config vllm`.",
+                "the model's `generation_config.json`" if src == "auto" else src,
+                str(diff_sampling_param),
+                scope="local",
             )
+
         return diff_sampling_param
 
     @property

vllm/entrypoints/openai/chat_completion/serving.py

@@ -143,14 +143,6 @@ class OpenAIServingChat(OpenAIServing):
         self.enable_prompt_tokens_details = enable_prompt_tokens_details
         self.enable_force_include_usage = enable_force_include_usage
         self.default_sampling_params = self.model_config.get_diff_sampling_param()
-        if self.default_sampling_params:
-            source = self.model_config.generation_config
-            source = "model" if source == "auto" else source
-            logger.info(
-                "Using default chat sampling params from %s: %s",
-                source,
-                self.default_sampling_params,
-            )
         if self.model_config.hf_config.model_type == "kimi_k2":
             self.tool_call_id_type = "kimi_k2"
         else:

vllm/entrypoints/openai/completion/serving.py

@@ -72,16 +72,9 @@ class OpenAIServingCompletion(OpenAIServing):
         self.logits_processors = self.model_config.logits_processors
 
         self.enable_prompt_tokens_details = enable_prompt_tokens_details
-        self.default_sampling_params = self.model_config.get_diff_sampling_param()
         self.enable_force_include_usage = enable_force_include_usage
-        if self.default_sampling_params:
-            source = self.model_config.generation_config
-            source = "model" if source == "auto" else source
-            logger.info(
-                "Using default completion sampling params from %s: %s",
-                source,
-                self.default_sampling_params,
-            )
+
+        self.default_sampling_params = self.model_config.get_diff_sampling_param()
 
     async def render_completion_request(
         self,

vllm/entrypoints/openai/responses/serving.py

@@ -221,15 +221,8 @@ class OpenAIServingResponses(OpenAIServing):
         )
         self.enable_prompt_tokens_details = enable_prompt_tokens_details
         self.enable_force_include_usage = enable_force_include_usage
+
         self.default_sampling_params = self.model_config.get_diff_sampling_param()
-        if self.default_sampling_params:
-            source = self.model_config.generation_config
-            source = "model" if source == "auto" else source
-            logger.info(
-                "Using default chat sampling params from %s: %s",
-                source,
-                self.default_sampling_params,
-            )
 
         # If False (default), the "store" option is (silently) ignored and the
         # response is not stored. If True, the response is stored in memory.